[BUG] scan_add_inclusive uses float division in range, triggering AST warning

[1ckendall]

Description

When using ti.algorithms.PrefixSumExecutor, the internal kernel scan_add_inclusive triggers a warning from the AST transformer:
Warning: Casting range_for boundary values from f32 to i32, which may cause numerical issues.

This appears to be caused by the use of the standard division operator / inside a range() call in the library code, which returns a float in Python 3.

Steps to Reproduce

Run the following minimal script on a GPU architecture (CUDA/Vulkan):

import taichi as ti

ti.init(arch=ti.gpu)

# Minimal size to trigger the kernel
N = 1024
arr = ti.field(ti.i32, shape=N)
arr.fill(1)

# Initialize executor
prefix_sum = ti.algorithms.PrefixSumExecutor(N)

# The warning appears during the first run when JIT compilation happens
prefix_sum.run(arr)
print("Finished")

Observed Behavior

The code runs, but prints the following warning during compilation:

[Taichi] version 1.7.4, ...
...
site-packages\taichi\lang\ast\ast_transformer.py:63: Warning: Casting range_for boundary values from f32 to i32, which may cause numerical issues
  warnings.warn(

Source of the Bug

The issue seems to be in taichi/_kernels.py inside the scan_add_inclusive function.

Specifically, this loop:

# Inter-warp scan, use the first thread in the first warp
if warp_id == 0 and lane_id == 0:
    for k in range(1, BLOCK_SZ / WARP_SZ):  # <--- CAUSE: returns float (e.g. 2.0)
        pad_shared[k] += pad_shared[k - 1]

Since BLOCK_SZ (64) and WARP_SZ (32) are integers, BLOCK_SZ / WARP_SZ returns 2.0 (float), which triggers the AST warning when passed to range.

Environment

• Taichi Version: 1.7.4
• Python Version: 3.12.10
• OS: Windows
• Arch: CUDA

Suggested Fix

Replace the division operator / with the integer division operator //:

# Change this:
for k in range(1, BLOCK_SZ / WARP_SZ):

# To this:
for k in range(1, BLOCK_SZ // WARP_SZ):